基于长时间视频序列的背景建模方法研究  

Background Modeling for Long-term Video Sequences

作　　者：丁洁[1,2] 肖江剑[2] 况立群[1] 宋康康[2] 彭成斌 DING Jie;XIAO Jiang-Jian;KUANG Li-Qun;SONG Kang-Kang;PENG Cheng-Bin(Computer and Control Engineering, North University of China, Taiyuan 030051;Computer Vision Group, Ningbo Institute of Industrial Technology, Chinese Academy of Sciences, Ningbo 315201)

机构地区：[1]中北大学计算机与控制工程学院,太原030051 [2]中国科学院宁波工业技术研究院计算机视觉团队,宁波315201

出　　处：《自动化学报》2018年第4期707-718,共12页Acta Automatica Sinica

基　　金：国家自然科学基金(61379080;61273276);浙江省杰出青年基金(LR13F020004);国家科技支撑计划(2015BAF14B01);钱江人才计划(QJD1702031);中国博士后科学基金(2017M612047)资助~~

摘　　要：针对现有背景建模算法难以处理场景非平稳变化的问题,提出一种基于长时间视频序列的背景建模方法.该方法包括训练、检索、更新三个主要步骤.在训练部分,首先将长时间视频分段剪辑并计算对应的背景图,然后通过图像降采样和降维找到背景描述子,并利用聚类算法对背景描述子进行分类,生成背景记忆字典.在检索部分,利用前景像素比例设计非平稳状态判断机制,如果发生非平稳变换,则计算原图描述子与背景字典中描述子之间的距离,距离最近的背景描述子对应的背景图片即为此时背景.在更新部分,利用前景像素比例设计更新判断机制,如果前景比例始终过大,则生成新背景,并更新背景字典以及背景图库.当出现非平稳变化时(如光线突变),本算法能够将背景模型恢复问题转化为背景检索问题,确保背景模型的稳定获得.将该框架与短时空域信息背景模型(以ViBe、MOG为例)融合,重点测试非平稳变化场景下的背景估计和运动目标检测结果.在多个视频序列上的测试结果表明,该框架可有效处理非平稳变化,有效改善目标检测效果,显著降低误检率.Considering the difficulties to deal with scene non-stationary variation of proposed background modeling methods, we propose a method for moving targets by exploiting periodic spatial-temporal feature from a long-term video.We use three steps, training, retrieval and updating, to establish a background modeling framework for long-term video sequences. In the training step, we cut hours of video into a number of minute clips and compute the average background to generate a series of background images. After performing resize and dimension reduction on background images, a set of descriptors are obtained for the clustering process, where background descriptors are classified into different clusters and each cluster is represented by a typical background image in the background memory dictionary. In the retrieval step, we use foreground pixel ratio as a criterion to determine sudden change of background. For those scenarios, the current image is converted to a background descriptor and compared to the descriptors stored in retrieval database to find a suitable background frame. If no similar background descriptor is found in the database, a new background image is to be generated and added into our dictionary and background image database. Using this framework, the background modeling problem is converted to a background retrieval problem when non-stationary change happens especially for the indoor scene with quick illumination changes such as light on/off. Combining the popular ViBe or MOG algorithm with our framework, we test a number of long term video sequences and achieve better results in terms of tracking targets and the false detection rate.

关 键 词：背景建模 长周期视频 背景图描述子 背景检索 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]